Search

The formula for scaling how much information in relations on the finite universe is proposed, which is called the entropy of relation R and denoted by H (R). Based on the concept of H (R), the entropy of predicates and the information of propositions are measured. We can use these measures to evaluate predicates and choose the most appropriate predicate for some given cartesian set. At last, H (R) is used to induce decision tree. The experiment show that the new induction algorithm denoted by IDIR do better than ID3 on the aspects of nodes and test time. [ABSTRACT FROM AUTHOR]

Behavioral sequences of the medaka (Oryzias latipes) were continuously investigated through an automatic image recognition system in increasing temperature from 25°C to 35°C. The observation of behavior through the movement tracking program showed many patterns of the medaka. After much observation, behavioral patterns could be divided into basically 4 patterns: active- smooth, active-shaking, inactive-smooth, and inactive-shaking. The "smooth" and "shaking" patterns were shown as normal movement behavior, while the "smooth" pattern was more frequently observed in increasing temperature (35° C) than the "shaking" pattern. Each pattern was classified using a devised decision tree after the feature choice. It provides a natural way to incorporate prior knowledge from human experts in fish behavior and contains the information in a logical expression tree. The main focus of this study was to determine whether the decision tree could be useful in interpreting and classifying behavior patterns of the medaka. [ABSTRACT FROM AUTHOR]

Axial representation plays a significant role in character recognition. The strokes of a character may consist of two regions, i.e. singular and regular regions. Therefore, a method to extract the central axis of a character requires two different processes to compute the axis in theses two different regions. The major problem of most traditional algorithms is that the extracted central axis in the singular region may be distorted by artifacts and branches. To overcome this problem, the wavelet-based amendment processing technique is developed to link the primary axis, so that the central axis in the singular region can be produced. Combining with our previously developed method for computing the primary axis in the regular region, we develop a novel scheme of extracting the central axis of character based on the wavelet transform (WT). Experimental results show that the final axis obtained from the proposed scheme closely resembles the human perceptions. It is applicable to both binary image and gray-level image as well. The axis representation is robust against noise. [ABSTRACT FROM AUTHOR]

Double parallel feedforward neural network (DPFNN) based approach is proposed for dimensionality reduction, which is one of very significant problems in multi- and hyperspectral image processing and is of high potential value in lunar and Mars exploration, new earth observation system, and biomedical engineering etc. Instead of using sequential search like most feature selection methods based on neural network (NN), the new approach adopts feature weighting strategy to cut down the computational cost significantly. DPFNN is trained by a mean square error function with regulation terms which can improve the generation performance and classification accuracy. Four experiments are carried out to assesses the performance of DPFNN selector for high-dimensional data classification. The first three experiments with the benchmark data sets are designed to make comparison between DPFNN selector and some NN based selectors. In the fourth experiment, hyperspectral data, that is an airborne visible/infrared imaging spectrometer (AVIRIS) data set, is used to compare DPFNN selector with widely used forward sequential search methods using the Maximum Likelihood classifier (MLC) as criterion. Experiments show the effectiveness of the new feature selection method based on DPFNNs. [ABSTRACT FROM AUTHOR]

A dimensionality reduction technique is presented for semi-supervised face recognition where image data are mapped into a low dimensional space with a spectral method. A mapping of learning data is generalized to a new datum which is classified in the low dimensional space with the nearest neighbor rule. The same generalization is also devised for regularized regression methods which work in the original space without dimensionality reduction. It is shown with experiments that the spectral mapping method outperforms the regularized regression. A modification scheme for data similarity matrices on the basis of label information and a simple selection rule for data to be labeled are also devised. [ABSTRACT FROM AUTHOR]